Hit the Antigravity weekly limit? The cheapest fallback

What happens when you hit the Antigravity weekly limit

The model you were using stops until its quota refreshes. On Google AI Pro, Google says the quota refreshes "every five hours until the weekly limit is reached", per Antigravity's plans page. Free and Google AI Plus get a weekly quota only. Ultra gets the highest five-hour quota and the highest weekly limit.

Google doesn't publish hard numbers. The quota tracks how much work the agent does, so a long refactor drains it faster than a string of quick edits. Your remaining quota per model is on the Settings page. See Antigravity limits explained for the plan-by-plan breakdown.

Switch to Claude in Antigravity first

This costs nothing extra and it's easy to miss. The model picker on Google's models page shows two separate sets of meters: one weekly and one five-hour limit for Gemini models, and another pair for Claude and GPT models. Running out of Gemini quota leaves the Claude meter untouched.

Claude Sonnet 5.5 and Claude Opus 5.5 are available on paid Google AI Pro (not trials) and on Ultra. Both rank above Gemini 4 Argon on Code Arena: Opus 5.5 is #1 of 50, Sonnet 5.5 is #3 of 50 and Argon is #8 of 50. Pick the model under the prompt box and keep going. A turn that's already running finishes on the old model.

One date to know: Google lists Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS-120b as "Will be removed on November 18, 2026." Those are the only Claude models on Free and Google AI Plus, so after Nov 18 the lower plans lose that escape hatch.

How AI credits work in Antigravity

Google AI Pro and Ultra can keep using a model after its quota runs out by spending AI credits, per Google. A setting called AI Credit Overages decides what happens: Never waits for the refresh, and Always spends credits and switches back to your plan quota once it refreshes. Free and Plus can't use credits.

The catch is price. Google says credits are consumed "at standard Gemini Enterprise consumption pricing" and, on its Google One help page, that "The amount of AI credits used for each feature and model varies." It doesn't publish a credit cost per token or per request on its public pages, so we can't give you a per-session figure. Check the AI credits activity page after a session to see what it took.

You also can't buy credits if you pay for Google One through Apple, Pixel Pass or another third party, and Antigravity credits aren't sold in Japan, per the same help page. If you want a predictable bill, leave overages on Never and use one of the paid options below.

Can you use your own Gemini API key in Antigravity?

No. Antigravity's plans page says there's no support for bring-your-own-key or bring-your-own-endpoint "for additional rate limits". Paying per token means using a different tool.

Gemini CLI with an API key: still works, with caveats

Gemini CLI is no longer the Google tool for AI Pro and Ultra. Its README on GitHub says free and Google AI Pro / Ultra accounts moved to Antigravity CLI on June 18, 2026, and that Gemini CLI isn't tuned for the Gemini 4 series. It's still maintained for Gemini Code Assist licenses and "legacy paid API / Vertex keys".

So you can still run it on a paid Gemini API key, which bills per token and doesn't touch your Antigravity quota. Set GEMINI_API_KEY to a key from Google AI Studio, run gemini and choose "Use Gemini API key", per the Gemini CLI auth docs. /stats model shows what a session used.

Gemini 3.8 Flash is the sensible model there at $0.75 in, $0.075 cached and $3.75 out per 1M tokens, about $3.98 for our sample session. It ranks #25 of 50 on Code Arena, though. For comparison, Gemini 4 Argon on the API costs $2 in and $10 out, about $9.80 a session. That's the launch price. From Dec 1, Google charges $4 in and $20 out, so the same session costs roughly twice as much. Prices are from Google's Gemini API pricing page.

The cheapest fallback overall: DeepSeek V4.1 Flash

DeepSeek V4.1 Flash ranks #12 of 50 on Code Arena, well above Gemini 3.8 Flash, and costs about $1.25 for our sample session at peak. Off-peak is half, per DeepSeek's pricing page, and peak hours fall in the US evening and early morning, so a normal US workday is off-peak. See DeepSeek off-peak hours.

It doesn't run inside Antigravity. DeepSeek documents setups for Claude Code and Codex instead, and we walk through both in Claude weekly limit hit: cheapest API fallback and Codex weekly limit hit: cheapest API fallback. Your code goes to DeepSeek under its terms, not Google's.

Should you upgrade to Google AI Ultra instead?

Only if you hit the weekly limit most weeks. Google says the $99.99 Ultra plan gives 5x the tokens of the $19.99 Pro plan and the $199.99 plan gives 20x, per its plan change post. The $80 gap between Pro and the first Ultra tier buys about 8 sample sessions on Gemini 4 Argon at today's API price, or about 64 on DeepSeek V4.1 Flash at peak (our math).

If you run out once a month, switch to Claude in Antigravity and stop there. If you run out every week and want to stay on Argon, Ultra is the cleaner fix. See is Gemini Ultra worth it and is Google AI Pro worth it.

Related comparisons

• Claude Sonnet 5.5 vs Gemini 4 Argon and Claude Opus 5.5 vs Gemini 4 Argon.
• DeepSeek V4.1 Flash vs Gemini 3.8 Flash and DeepSeek V4 Pro vs Gemini 4 Argon.
• Claude vs Gemini API pricing and the cheapest LLM APIs.